Interpolation
Filling the space between two points with several in-between steps
- Interpolation sets two endpoints and fills the space between them with several steps.
- What actually gets blended isn't the finished result — it's the coordinates that carry its meaning. That's very different from overlaying two pictures translucently.
- So every in-between step is a whole, complete result in its own right — not a half-see-through ghost, but a real picture or melody doing its own job.
- In video it smooths the gap between two scenes; in music it lets one melody ease naturally into another.
- The path isn't always smooth. A dead stretch or a sudden jump can show up along the way.
Contents
1The analogy
Squeeze blue paint on one side of a palette, yellow on the other, and divide the space between into five wells. Mix a slightly different ratio in each well and you get a strip running from blue through green to yellow. The middle well isn't a blend of blue and yellow showing through — it's a complete, solid green.
That mixing is exactly the bridge interpolation builds between two results. Set a starting point and an ending point, decide how many wells to divide the space into, and out comes that many in-between results.
That's different from laying blue and yellow paper on top of each other translucently. Stacked paper still shows both colors separately underneath; mixed paint is one new color. The in-between steps interpolation produces aren't two pictures showing through each other either — each one is a complete picture in its own right.
2In detail
What gets mixed is coordinates, not the finished result
A generative model never works with a picture or sound directly. It first turns whatever meaning is inside into a bundle of numbers, then builds the result from those numbers. The space those number bundles live in is called latent space.
Interpolation happens on those coordinates. Turn two pictures each into their own coordinates, pick several points along the line between them, and unfold each point back into a picture. Since one coordinate becomes one whole picture, every in-between point becomes a whole picture too.
Blend the pixels of two pictures directly and you get something very different — half of one picture and half the other layered on top, looking like a ghostly double exposure. The two methods produce pictures that look nothing alike.
How many steps to divide it into
The step count is up to you. Split it into three and it feels choppy; split it into thirty and it flows like water. Video work usually reaches for a few dozen steps.
Even a fine split doesn't spread the change evenly. Some stretch stays nearly unchanged while another stretch shifts sharply — a smile can suddenly appear at the corners of the mouth in just one step, rather than easing in gradually.
That's why it's common to lay the results out, spot the stretch that changes too fast, and rebuild just that stretch with a finer split. It's less work than making the whole thing finer across the board.
What it does in video and music
In video, it fills the gap between one scene and the next. A video with too few frames per second looks choppy; slotting interpolated frames in between smooths it out. It works out what moved where between the frames on either side, then builds the spot in between.
In music, it bridges two melodies. Set a calm melody and an upbeat one as the two ends, fill the space between, and you get a stretch that eases gradually from one into the other — useful for transitioning between tracks or building a looping background piece.
It's used for camera movement too. Set just a starting position and an ending position and it fills in a smooth path between them.
The path isn't always smooth
There's no guarantee that every point along the line between two coordinates makes sense. Pass through a stretch the training material barely covered, and the shape falls apart into something strange. The more different the two endpoints are, the more likely the path runs through a stretch like that.
How the direction is handled changes the result too. A straight line between two coordinates tends to dip and blur in the middle, so some tools instead follow an arc, like tracing along a sphere. The same two endpoints can produce a different middle depending on which path you take.
Set apart from variation and consistency
Variation pulls out several sibling results from the same prompt, side by side, with no order to them. Interpolation has a fixed start and end, lined up in a single sequence — swap the first and last step and the direction reverses.
Consistency means holding the same subject steady across several results. Interpolation does the opposite on purpose, showing a gradual change happening — the two pull in different directions entirely.
3More precisely
Interpolation is a calculation that picks points along the line connecting two coordinates in latent space and unfolds each one back into a result. Connecting two points with a straight line is called linear interpolation; following an arc along a sphere is also widely used. Text-based tools sometimes blend the number bundle that encodes a prompt, or the starting noise, instead of coordinates — what gets blended changes the character of the in-between results.
The paint analogy breaks down in a few places. Mix paint in any ratio and you always get a color, but the space between two coordinates can include stretches that don't mean anything at all. Mixing two paints also lets you roughly predict the result just from the two colors, while the middle of two coordinates has to actually be built to know what it looks like. A palette gives each color one spot, while latent space has hundreds or thousands of directions, so even the same two endpoints can be connected by several different paths. Which path looks most natural is hard to know ahead of time — building a few and comparing them by eye is usually the fastest way to pick.
4Try it yourself
5Common misconceptions
It's easy to think this is the same as overlaying two pictures translucently, but actually it blends coordinates, so every in-between step is a complete picture with no overlap showing through.
It's easy to think a finer split spreads the change out evenly, but actually some stretches barely move while others shift sharply.
It's easy to think picking any two results guarantees a smooth path between them, but actually the shape can fall apart in the middle stretch.
7One-line summary
In shortInterpolation divides the space between two points into several steps on the coordinates that carry meaning, not the finished result, so every step comes out as a complete result of its own.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02